Breaker: Removing Shortcut Cues with User Clustering for Single-slot Recommendation System

Chao Wang, Yupeng Zheng, Yujing Zhang, Yan Feng, Zhe Wang, Xiaowei Shi, An You, Yu Chen

Abstract

In single-slot recommendation systems, users are exposed to only one item at a time, preventing the simultaneous collection of feedback on multiple items. This constraint necessitates pointwise modeling, which focuses on individual user-item interactions rather than direct ranking. However, because user-side data is often significantly richer than item-side data, models frequently learn "intrinsic user tendencies" as a shortcut bias, neglecting nuanced user-item preferences. We propose Breaker, a framework that integrates an auxiliary user representation clustering task with a multi-tower structure for cluster-specific preference modeling. By grouping users with similar characteristics, Breaker increases the complexity of the pointwise task, forcing the model to better capture item-specific preferences. Deployed at Meituan, Breaker serves tens of millions of users daily, demonstrating superior performance in both offline and online evaluations.

The Challenge: Shortcut Bias in Single-Slot Systems

Single-slot recommendation scenarios, such as splash screen advertisements or targeted marketing slots, present a unique challenge. Unlike list-based feeds where a user's choice between items reveals ranking information, single-slot systems only observe a binary outcome: did the user interact with the single item shown?

Figure 1: (a) Examples of single-slot recommendation interfaces. (b) Feature importance analysis showing user features (U_1-U_5) dominating item features (I_1). (c) Minimal performance gain when adding item-related features to standard models.

As shown in Figure 1, standard deep neural networks (DNNs) often take a "shortcut" by relying on rich user-side information (e.g., historical click-through rates, demographics) to predict engagement. This leads the model to identify "frequent clickers" rather than understanding why a user prefers a specific item. Consequently, adding item-side features often yields negligible performance improvements because the model has already achieved low loss by simply modeling user-intrinsic tendencies.

Methodology: The Breaker Framework

To eliminate these shortcut cues, the Breaker model introduces a multi-stage architecture designed to "force" the model to attend to item features:

  1. Representation Extraction Module (REM): Generates latent embeddings for users and items using standard multi-layer perceptrons (MLPs).
  2. User Representation Clustering Module (URCM): Employs Deep Embedding Clustering (DEC) to group users into $K$ clusters. By clustering users with similar intrinsic tendencies, the model can no longer rely on those tendencies to differentiate users within the same cluster.
  3. Cluster-specific Preference Modeling Module (CPMM): A multi-tower architecture where each tower learns preferences for a specific user cluster. The final prediction is a weighted sum of tower outputs based on cluster membership probabilities.

Training Stability: Delayed Parameter Updates

Jointly training unsupervised clustering with supervised classification is inherently unstable. To address this, Breaker utilizes a delayed parameter update mechanism (inspired by target networks in Reinforcement Learning). A target network for clustering is maintained and synchronized with the main network only every $M$ steps, ensuring stable cluster assignments and smoother convergence.

Figure 6: The overall system architecture at Meituan, illustrating the integration of the Campaign Configuration Platform, the Algorithm Platform (where Breaker resides), and the Data Development Platform.

Experimental Results

Breaker was evaluated on two large-scale datasets from Meituan, collected via Randomized Controlled Trials (RCT) to ensure unbiased evaluation.

Offline Evaluation

In high-engagement scenarios, Breaker achieved a 7.05% improvement in Recall@1 and a 5.65% improvement in Average Expected Response (AER) over the strongest baselines (including LightGBM, DeepFM, and DCN-V2). A key insight was that as training progressed, the model's ability to rank intrinsic user tendencies (item-based AUC) decreased while its ability to select the correct item (Recall@1) increased, confirming the successful removal of shortcut bias.

Online Impact

During online A/B testing at Meituan, Breaker delivered significant business impact:

  • Scenario A (High-engagement): 4.93% lift in conversion rate (CVR) compared to DeepFM.
  • Scenario B (Sparse interaction): 7.15% lift in CVR compared to LightGBM.
  • Efficiency: The model is approximately 40% faster during inference than complex neural baselines like xDeepFM, making it highly suitable for high-traffic production environments.

Conclusion

Breaker effectively addresses the shortcut bias inherent in single-slot recommendation systems by leveraging unsupervised user clustering to increase task complexity. This forces the model to move beyond general user tendencies and capture genuine user-item preferences. Future work will explore dynamic clustering to account for evolving user behaviors over time.

Summary updated

Abstract

In a single-slot recommendation system, users are only exposed to one item at a time, and the system cannot collect user feedback on multiple items simultaneously. Therefore, only pointwise modeling solutions can be adopted, focusing solely on modeling the likelihood of clicks or conversions for items by users to learn user-item preferences, without the ability to capture the ranking information among different items directly. However, since user-side information is often much more abundant than item-side information, the model can quickly learn the differences in user intrinsic tendencies, which are independent of the items they are exposed to. This can cause these intrinsic tendencies to become a shortcut bias for the model, leading to insufficient mining of the most concerned user-item preferences. To solve this challenge, we introduce the Breaker model. Breaker integrates an auxiliary task of user representation clustering with a multi-tower structure for cluster-specific preference modeling. By clustering user representations, we ensure that users within each cluster exhibit similar characteristics, which increases the complexity of the pointwise recommendation task on the user side. This forces the multi-tower structure with cluster-driven parameter learning to better model user-item preferences, ultimately eliminating shortcut biases related to user intrinsic tendencies. In terms of training, we propose a delayed parameter update mechanism to enhance training stability and convergence, enabling end-to-end joint training of the auxiliary clustering and classification tasks. Both offline and online experiments demonstrate that our method surpasses the baselines. It has already been deployed and is actively serving tens of millions of users daily on Meituan, one of the most popular e-commerce platforms for services.
BibTeX
@misc{wang2025breaker,
  author = {Chao Wang and Yupeng Zheng and Yujing Zhang and Yan Feng and Zhe Wang and Xiaowei Shi and An You and Yu Chen},
  title = {{Breaker: Removing Shortcut Cues with User Clustering for Single-slot Recommendation System}},
  year = {2025},
  eprint = {2506.00828},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.00828},
  howpublished = {arXiv},
}
Figure